Hi @AntonioDeDomenico, Thanks for the clarification on (2)! Just to make sure I understand the intended scope of preprocessing: if larger models are allowed in preprocessing, are there any constraints on using them for semantic reasoning steps (e.g., decomposition, chain-of-thought generation, or intermediate solution drafting), as opposed to purely mechanical steps like retrieval, filtering, or formatting?

I’m asking because, depending on how preprocessing is defined, if using Larger models at inference time is allowed for preprocessing , the problem can in turn be divided into two stages where reasoning can be heavily offloaded to larger model like generating CoT and use the track model only for final summarization, which could blur the distinction between compute-constrained tracks.

Thankyou for your prompt replies and clarifications!!

Hi @ahuvam, @neuron_x indeed, this is too blurry, and I should have been more precise. Larger models should not be used during inference time. I think this is clear enough and do not let space to misunderstanding

yes, thanks a lot!!